How blind obedience shaped one of history’s most unsettling experiments, and why its lessons matter more than ever in the age of AI

In August of 1961, Stanley Milgram, a psychologist at Yale University, set out to test how far ordinary people would go when instructed by an authority figure to act against their own conscience. This was a particularly important experiment at the time: three months earlier, the televised trial of Nazi war criminal Adolf Eichmann had started in Jerusalem. In later writings, Milgram said that he designed his experiment because he was fascinated by Eichmann’s claim (similar to the claims made by other members of the Nazi party) that he was “simply following orders”.

There were 40 participants in the original experiment, who were all kept in the dark about its true nature. Instead, they were told that they were assisting in a study on memory and learning. Their task as the “educator” in the experiment was to teach an unseen adult “learner” a list of word pairs, and then the learner had to repeat the word pairs that they could remember. When a learner gave a wrong answer, the educator was expected to administer an electric shock to the learner. With each mistake, the voltage increased, from 15 volts all the way up to 450, levels clearly marked from “Slight Shock” to “Danger: Severe Shock.” Unbeknownst to the educator, the shocks were not real. Had they been real, however, the highest setting could have been fatal.

Before the experiment began, Milgram asked fourteen senior psychology students to predict the outcome. They estimated that perhaps one or two people out of a hundred would deliver the maximum voltage. His colleagues in the psychology department agreed with this hypothesis. The expectation was that conscience would prevail, and that participants would simply refuse to administer shocks above a certain level.

But surprisingly, that wasn’t the outcome of the experiment at all. 

The lab coat and the switch

Each session involved three people: the experimenter, dressed in a lab coat; the volunteer educator, who believed they were assisting the experimenter; and the learner, who was in fact an actor working with the research team.

The learner was introduced to the educator at the beginning of the session and then strapped into what appeared to be an electric chair. Before the test began, the educator received a sample shock to make the setup believable. The educator was then moved into an adjacent room where they could not see the learner, but they could hear him. They started reading out the word pairs so that the learner could memorise them and recite them back. When answers were wrong, the educator pressed a switch and the learner “reacted” to the fake shock they received. As the voltage climbed with each wrong answer, the learner protested more and more loudly. He pleaded and cried. He mentioned a heart condition. And then, after the highest level of shock was administered, he fell ominously silent.

Most of the learner’s cries were actually prerecorded. But of course the educators didn’t know that. They were led to believe that they were inflicting real pain on a real human being. When participants hesitated to administer shocks, the experimenter (who was also in the room) responded with a pre-planned series of carefully worded, escalating prods:

“Please continue.”

“The experiment requires that you continue.”

“It is absolutely essential that you continue.”

“You have no other choice; you must go on.”

If, after all four prompts, a participant still refused to administer a shock, the session ended. Otherwise, it continued until the maximum voltage had been administered three times. Many participants showed visible distress during the process. They sweated, trembled, and stuttered as they attempted to teach the word pairs to the learner. Some dug their fingernails into their skin. 14 of the 40 laughed nervously. Every participant paused at least once to question what they were doing, but most continued after a verbal prod from the man in the lab coat. 

Milgram’s hypothesis had been that very few participants would be willing to follow the command to keep administering higher voltage shocks. In reality, every participant in the experiment, despite their obvious discomfort, continued to at least 300 volts. 65% of participants made it all the way past the learner’s desperate cries about his heart condition and still went up to the “fatal” 450 volts.

Those few who refused to go further did not storm out or demand that the study be stopped. They simply stopped pressing the switch. They did not stop to check on the learner on their way out or enquire about his condition. They simply left. 

Not a German problem

Before the experiment began, Milgram believed that he already knew what he was testing for. He suspected that the obedience displayed by Nazi perpetrators during the Holocaust reflected something culturally specific, perhaps a distinctly German disposition toward authority. American participants, he assumed, would serve as a kind of moral control group. Later, he planned to replicate the study in Germany, expecting far higher levels of compliance there. The results of his experiment made that second phase unnecessary.

The obedience Milgram observed wasn’t uniquely German. It turned out to be universal. When the experiment was later repeated in countries across the world, the pattern held. Different languages, cultures, and political histories, yet the same unsettling outcome.

Reflecting on his findings, Milgram wrote that the most disturbing discovery was not cruelty or hatred, but willingness: the readiness of ordinary adults to go to extreme lengths when instructed by an authority figure. He knew that these people weren’t acting out of malice – he could see that in the way that they were unsettled and upset by their actions. They were simply doing what they believed their role required of them, and in so doing, they became participants in a deeply destructive process.

To explain this, Milgram proposed what he called the agentic state. In this state, individuals no longer see themselves as autonomous moral actors. Instead, they begin to view themselves as instruments carrying out the wishes of someone else. Responsibility shifts upward and rests with the authority. Once that psychological handover occurs, obedience becomes easier, doubt becomes easier to squash, and personal conscience fades into the background.

Obedient machines

Human beings are having a lot of important conversations about artificial intelligence at the moment, and many of those conversations follow a familiar cadence: alarm, outrage, and a deep unease about where this technology is taking us. Headlines regularly spotlight the harms of AI models churning out fake or explicit content, bots trained on stolen images or copyrighted works, and automated systems amplifying misinformation with unprecedented reach. In one widely reported case, users of the AI chatbot Grok were able to generate sexually explicit deepfake images of women and children without their consent – hundreds of them in mere hours – prompting global backlash and regulatory scrutiny across Europe, Asia, and the Americas.

These incidents feel, to many, like evidence that AI has become a rogue force; something that needs to be restrained, feared, or outright banned. But beneath the surface of each controversy lies the thorny truth: AI does not choose to harm. It follows commands – the prompts we give it, the frameworks we build, the incentives we embed. In that sense, when an AI spits out harmful deepfakes or spin-doctor political content, it is not “choosing evil” any more than the educator in Milgram’s lab coat was choosing cruelty. It is responding to human input, just as Milgram’s subjects responded to his authority.

If AI misuse feels so threatening, it’s because the scale and speed at which AI operates magnify the consequences of bad orders. A malicious prompt can generate thousands of damaging images in an hour. A misleading text snippet can be amplified across social platforms before fact-checkers have a chance to blink. Governments, watchdogs, and civil society groups are scrambling to catch up, drafting laws and safety guidelines designed to curb these harms and protect individuals and communities.

But there is a deeper philosophical question here, one that echoes the lessons of Milgram’s research: is the technology itself the problem, or the human willingness to misuse it?

What if we could encode ethical constraints into the code in a way that prevents harmful obedience altogether? What if the “obedient agentic state” (the prediscussed tendency to follow orders without moral resistance) could be designed out of our machines? Some scholars argue that defining ethical AI isn’t just about regulation or supervision, but about embedding moral reasoning and accountability into the systems themselves. The field of machine ethics explores whether AI can be given frameworks that go beyond blind rule-following, allowing it to evaluate the impact of a request rather than merely executing it.

That prospect raises difficult but important questions. If we could design AI that refuses harmful instructions, such as the instruction to generate content that violates privacy, dignity, or consent, could such systems ultimately behave more ethically than the average human? In theory, yes. An AI’s actions could be governed not by expediency or obedience to authority, but by consistent ethical principles rigorously tested against edge cases humans frequently misjudge. In practice, this would require designers to prioritise ethics over profit, and legislators to enforce accountability with teeth – no small task, given how quickly the technology evolves.

In this light, the ethical promise of AI isn’t about fearing its autonomy. It’s about confronting our own. After all, we are the ones issuing the orders. And if we want AI to reflect our highest ideals – our respect for each other’s rights, dignity, and wellbeing – then the first step is to recognise that technology mirrors human intent, for better or worse.

About the author: Dominique Olivier

Dominique Olivier is the founder of human.writer, where she uses her love of storytelling and ideation to help brands solve problems.

She is a weekly columnist in Ghost Mail and collaborates with The Finance Ghost on Ghost Mail Weekender, a Sunday publication designed to help you be more interesting. She now also writes a regular column for Daily Maverick.

Dominique can be reached on LinkedIn here.